MCP Server
MCP tools for guardrails, approvals and agent activity
The tools Claude, ChatGPT and other MCP clients use to follow a project’s guardrails, learn from feedback, go through approvals and report what they did, plus the messages they get when something is refused.
These tools are available to any MCP client connected to Reaudit (and to the desktop agent and Reaudit assistant). Ask in plain language, for example “check this post against our guardrails”, “what did you publish this week?” or “submit these posts for approval”.
Guardrails
get_project_guardrails: the project’s approved claims, banned phrases, disclosures and brand rules (with ids), the enforcement level, and what agents may change.check_draft_guardrails: check a draft for banned phrases, missing disclosures and (by default) factual claims against approved claims and the knowledge base. A project with no rules returns nothing to fix.update_project_guardrails: add or remove rules (by id or exact text), change enforcement or sensitive claim types, withdryRunto preview. Owners and admins only; Claude and ChatGPT ask you to confirm first.
Feedback, approvals and activity
list_agent_feedback/resolve_agent_feedback: drafts people rewrote or deleted, and saving the lesson (brand rule, banned phrase or approved claim). Only when Learn from feedback is on.submit_for_approval/get_approval_status: send drafts through the approval chain and check where they are. Agents cannot approve.list_agent_activity: what agents changed on the project, newest first, including blocked publishes.get_brand_examples: approved, team-rewritten and published examples of the brand’s work, plus brand images, to match before writing.propagate_fact_change: preview (default) or apply a fact change across drafts, posts, ads, the knowledge base and guardrails.get_creative_breakdown/rebuild_creative_breakdown: which ad hooks, offers, calls to action and images performed above average.
Paid ads tools now work over MCP
Platform staff can also use the admin news monitor and CRM advocacy tools from MCP clients again. Ad campaigns, ad creatives, connected ad accounts, visibility lift, campaign briefs from visibility gaps and the Paid Intelligence reports and actions (search terms, keyword and ad performance, Meta audience and creative insights, bids, negative keywords and switching Meta campaigns on or off) can now be used from Claude, ChatGPT and other MCP clients. They follow the same rules as the dashboard, and switching a campaign on goes through the project’s guardrails and approvals first.
When something is refused
guardrail_blocked: a publish broke the project’s guardrails (enforcement “Block”). The message lists what to fix.approval_required: the approval chain is on and the item is not approved in its current form. Submit it, or wait for the current step.agent_permission_denied: this kind of agent may only prepare drafts, or only read, on this project.agent_edit_not_allowed: the project does not let agents make that change to its guardrails.- Every refusal is written to Agent activity with the reason.
Background: [Keeping AI agents in check](/dashboard/help?article=agent-control-overview).